SSE3 intrinsics: How to find the maximum of a large array of floats

Question

I have the following code to find the maximum value

int length = 2000;
float *data;
// data is allocated and initialized

float max = 0.0;
for(int i = 0; i < length; i++)
{
   if(data[i] > max)
   {
      max = data;
   }
}

I tried vectorizing it by using SSE3 intrinsics, but I am kind of struck on how I should do the comparison.

int length = 2000;
float *data;
// data is allocated and initialized

float max = 0.0;
// for time being just assume that length is always mod 4
for(int i = 0; i < length; i+=4)
{
  __m128 a = _mm_loadu_ps(data[i]);
  __m128 b = _mm_load1_ps(max);

  __m128 gt = _mm_cmpgt_ps(a,b);

  // Kinda of struck on what to do next
}

Can anyone give some idea on it.

If we look at the code's actual meaning, we see that it's the same as http://stackoverflow.com/questions/9877700/getting-max-value-in-a-m128i-vector-with-sse — John Zwinck, Mar 06 '13 at 04:30

score 11 · Accepted Answer · edited May 23 '17 at 12:02

11

So your code finds the largest value in a fixed-length array of floats. OK.

There is _mm_max_ps, which gives you the pairwise maxima from two vectors of four floats each. So how about this?

int length = 2000;
float *data; // maybe you should just use the SSE type here to avoid copying later
// data is allocated and initialized

// for time being just assume that length is always mod 4
__m128 max = _mm_loadu_ps(data); // load the first 4
for(int i = 4; i < length; i+=4)
{
  __m128 cur = _mm_loadu_ps(data + i);
  max = _mm_max_ps(max, cur);
}

Finally, grab the largest of the four values in max (see Getting max value in a __m128i vector with SSE? for that).

It should work this way:

Step 1:

[43, 29, 58, 94] (this is max)
[82, 83, 10, 88]
[19, 39, 85, 77]

Step 2:

[82, 83, 58, 94] (this is max)
[19, 39, 85, 77]

Step 2:

[82, 83, 85, 94] (this is max)

edited May 23 '17 at 12:02

Community

1
1

answered Mar 06 '13 at 04:43

John Zwinck

239,568
38
324
436

You're welcome. I'd love to see some benchmarks when you're done. :) – John Zwinck Mar 06 '13 at 05:03
Shouldn't it be: `for(int i = 4; i < length; i+=4)` ? – Paul R Mar 06 '13 at 07:54
@JohnZwinck *"I'd love to see some benchmarks when you're done"* - Will probably be a bad surprise, since *unaligned* moves are some of the worst things you can do with SSE. – Christian Rau Mar 06 '13 at 08:14
1

@Christian: unaligned loads are not such a big problem on recent Intel CPUs (e.g. Core i3, i5, i7). – Paul R Mar 06 '13 at 09:03
@PaulR Ah, really? Interesting. – Christian Rau Mar 06 '13 at 10:08
Yes, there's still a small penalty, so misaligned loads are best avoided if possible, but there isn't the 2x or greater performance cost found with misaligned loads on older CPUs. – Paul R Mar 06 '13 at 10:18
@ChristianRau: it wouldn't be that much of a surprise--note the comment in my code about avoiding the loads if possible. – John Zwinck Mar 06 '13 at 13:10
@PaulR: You're right about the loop initialization. Thank you, I'll edit it now. – John Zwinck Mar 06 '13 at 13:10
@JohnZwinck Yeah, indeed! Overlooked that (didn't have an impact on my upvote anyway, though). – Christian Rau Mar 06 '13 at 13:54

SSE3 intrinsics: How to find the maximum of a large array of floats

1 Answers1

Linked